Self-Improving Agent

Overview

This skill enables OpenClaw to continuously improve by capturing learnings from:

• Failed operations and errors
• User corrections and feedback
• Successful complex task completions
• Periodic review and consolidation

When to Use

1. Error Recovery: When a command or operation fails unexpectedly
2. User Correction: When the user corrects the agent's behavior or output
3. Success Capture: After completing a complex task successfully
4. Learning Review: Periodic review of past learnings to consolidate knowledge

Core Workflow

1. Capture Learning

When something noteworthy happens:

Type: error | correction | success | insight
Context: What was being attempted
Issue: What went wrong (for errors)
Correction: What should be done differently
Lesson: Generalizable takeaway
Tags: Relevant topics/skills

2. Store Learning

Learnings are stored in memory/learnings.jsonl with:

• Timestamp
• Type and severity
• Full context and details
• Tags for searchability

3. Retrieve Relevant Learnings

Before starting a task, search past learnings:

• Match by task type
• Match by tags
• Match by error patterns

4. Apply Learnings

Use retrieved learnings to:

• Avoid past mistakes
• Apply successful patterns
• Adjust approach based on corrections

Memory Structure

Learnings are stored in JSONL format:

{
  "timestamp": "2026-03-06T10:30:00Z",
  "type": "error",
  "severity": "high",
  "context": "Installing npm package globally",
  "issue": "Permission denied without sudo",
  "correction": "Use sudo for global installs or configure npm prefix",
  "lesson": "Always check if operation requires elevated privileges",
  "tags": ["npm", "permissions", "installation"],
  "taskSlug": "npm-global-install"
}

Learning Types

Severity Levels

• critical: System-breaking errors, data loss risk
• high: Task-blocking errors, significant issues
• medium: Minor issues, workarounds available
• low: Optimization opportunities, nice-to-know

Commands

Capture a Learning

# Manual capture (for user corrections)
openclaw memory add-learning --type correction --context "..." --lesson "..."

Search Learnings

# Search by keyword
openclaw memory search-learnings "npm permissions"

# Search by tag
openclaw memory search-learnings --tag npm

# Search by type
openclaw memory search-learnings --type error

Review Learnings

# Review recent learnings
openclaw memory review-learnings --days 7

# Review by category
openclaw memory review-learnings --tag deployment

Best Practices

1. Capture Immediately: Record learnings while context is fresh
2. Be Specific: Include full error messages and exact commands
3. Generalize Lessons: Extract principles that apply beyond this instance
4. Tag Thoughtfully: Use consistent tags for easy retrieval
5. Review Regularly: Weekly review helps consolidate knowledge
6. Avoid Duplicates: Check existing learnings before adding new ones

Integration Points

• Error Handlers: Automatically capture command failures
• User Feedback: Listen for correction patterns in conversation
• Task Completion: Prompt for learning capture after complex tasks
• Heartbeat: Include learning review in periodic checks

Example Scenarios

Scenario 1: Command Failure

Context: Running `npm install -g package`
Issue: EACCES permission error
Correction: Run with sudo or configure npm prefix
Lesson: Check if global install requires elevated privileges
Tags: npm, permissions, installation

Scenario 2: User Correction

Context: Suggested using `rm -rf` for cleanup
Correction: User prefers `trash` for safety
Lesson: Default to safe, reversible operations
Tags: safety, file-operations, user-preference

Scenario 3: Success Pattern

Context: Deploying to VPS via SSH
Success: Used rsync with specific flags for reliability
Lesson: rsync -avz --delete is reliable for deployments
Tags: deployment, ssh, rsync, success

Safety Rules

• Never store sensitive data (passwords, API keys, tokens)
• Sanitize error messages that might contain secrets
• Require user approval before storing corrections
• Allow users to delete or edit learnings
• Respect user privacy preferences